acceptodds
Under review as a conference paper at ICLR 2027

Wired to Reason: Connectome Reservoirs Align with Vision-Language Models on Raven's Matrices

Abstract

The human brain and large neural networks have independently converged on surprisingly similar internal representations, a finding typically attributed to learning at scale. But what if a part of this convergence requires no gradient learning in the recurrent operators? We test this using a connectome-wired reservoir: a recurrent dynamical system whose bilinear operators are derived entirely from empirical brain connectivity matrices, with no learning in the recurrent weights. Remark- ably, the internal states this system builds during visual reasoning align with the layer-wise representations of a trained vision-language model in a structured and interpretable way: alignment varies systematically across recurrent states and VLM depth, with stronger alignment for correct trials in several model–metric combinations, without gradient-based optimization. This alignment is not a statistical artifact: empirical FC exhibits modest but consistently stronger alignment than the matched random control, while temporal-memory removal alters the alignment structure. Moreover, reservoirs instantiated with connectomes from both neurotypical and autistic individuals produce consistent visual reasoning performance that substantially exceeds random-weight baselines, indicating that this advantage generalizes across both cohorts rather than being tied to any particular population or diagnostic profile; structure-matched null controls further indicate that the empiri- cal connectome’s regional arrangement, not any single isolated property, underlies the alignment advantage. Taken together, our results suggest that the partial convergence between biological and vision-language models (VLMs) is not exclusively learned: FC-derived recurrent organization provides a structural inductive bias for partial CWR–VLM representational correspondence. Here, connectome-wired refers specifically to the use of an empirical brain connectivity matrix to determine the fixed recurrent dynamics of the reservoir. No measured neural activity is used as a representation target: all alignment analyses compare reservoir states with representations from trained vision-language models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Reservoir computing, connectome, VLMs, visual reasoning, alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_W7AiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper derives fixed bilinear reservoir operators from functional connectomes, trains a RAVEN readout, and compares three processing stages with VLM representations using CKA/RSA. It claims structured alignment, correctness selectivity and biological specificity. Macro accuracy is 61.96% versus 57.37% for uniform-random wiring, with extensive structured and preprocessing controls in the appendices.

Critical strengths

S1. The control suite includes rewiring, ROI permutation, structured graphs and preprocessing ablations. Controls matching or exceeding accuracy are disclosed alongside remaining alignment differences (Table 13).

S2. CWR-TD exceeds vectorized reservoir baselines on all seven tasks (Table 8). Random-wired CWR also performs well, supporting an architectural contribution independently of biological attribution.

Critical weaknesses

W1. Central interpretations conflict with the displayed figures. Fig. 2's Qwen fusion RSA is H1<H3<H2H_1<H_3<H_2, not the stated H1<H2<H3H_1<H_2<H_3; the Phi image title names 3x3Grid while its caption says Center. Fig. 3 shows higher raw CKA for error candidates on Center, contrary to its accompanying narrative, though this alone does not refute the reported across-task aggregate. Identical H1/H2H_1/H_2 states do not establish an unconfounded H3H_3-only test if VLM targets or instances differ. Random RSA in Fig. 4 and Null-Qwen CKA in Fig. 22a remain structured, making descriptions of undifferentiated or abolished alignment too strong. Nonzero raw CKA itself still needs a calibrated null reference.

W2. Biological specificity is not isolated. Each principal wiring matrix is fixed across seeds (App. K.4), and sparsity is selected using the biological-random gap (Table 3). Structured controls match or exceed accuracy while retaining lower CKA, and engineered signs outperform biological sign placement. These findings constrain, rather than eliminate, a connectome effect. ROI permutation also preserves the unlabeled graph while changing an order-sensitive preprocessing pipeline, so its effect does not alone identify a biological mechanism.

Questions for authors
  1. Figures: Please provide canonical stage labels and values for Figs. 2-4, resolve the Phi task mismatch, and reconcile the prose with the plotted ordering and residual alignment.
  2. Correctness: How are instances, distractors and VLM targets paired across conditions? Please give per-task correct-minus-error contrasts with uncertainty and justify localization to H3H_3.
  3. Controls: In a focused setting, where does the empirical wiring rank among independently sampled nulls using identical preprocessing and independently fixed hyperparameters? Include shuffled-pairing or input-only alignment references.
Reviewer_TdtrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work studies the representational alignment between FC constrained reservoir and VLM in visual reasoning. The paper use brain derived FC to define fixed recurrence and investigate how reservoir states align with VLM representations in reasoning stages. The work is evaluated with RAVEN and comparisons of reasoning performances and alignment against random and structured connectivity controls are included.

Critical strengths

The paper explores an interesting perspective with brain connectivity as in inductive biased with fixed recurrent weights; the evaluation is extensive, includes several VLMs, and includes both random and structured connectivity comparisons.

Critical weaknesses
  1. The biological interpretation of the connectivity is not fully established. First, it is resting state functional connectivity without sufficient justification. Second, the connectivity is processed with transformations such as random sign assignment, hence it is unclear what biological properties are retained and targeted. Additionally, authors state comparable reasoning performance between connectome reservoir and structured controls. Although stronger representational alignment is reported, it is not clear what property of the connectome wiring is responsible for this advantage.
  2. Authors states “removing temporal integration (alpha = 1)” at line 368; according to eq 2, this removes leaky integration, but the recurrent memory is present from Ht-1 (no leak vs no memory). Can the authors clarify the interpretation this ablation?
  3. The representational alignment results needs more careful interpretations. It appears H1, H2, and H3 are assigned different reasoning roles, while the states receive sequential input, unclear how they reflect different levels of reasoning abstraction. H3 receiving candidate specific information while H1 and H2 do not, may partially explain the alignment differences. Empirical connectivity were shown to have generally stronger CKA than controls, this does not necessarily establish that these states represent distinct reasoning stages. Could authors clarify evidence supporting this interpretation?
Questions for authors
  1. Prior connectome reservoir frameworks predominantly SC as the recurrent connectivity matrix, and this papers uses FC (temporal correspondence of BOLD). Could authors discuss how their approach is positioned relative to existing approaches and clarify the biological interpretation of using FC?
  2. With the above in mind, could the authors clarify what biological property is then retained and how do we interpret the alignment?
Reviewer_VhNqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds fixed bilinear reservoir operators from ABIDE functional-connectivity matrices and trains a downstream readout for RAVEN problems. It compares three sequential reservoir states with layer representations of task-fine-tuned Qwen2-VL-2B and Phi-3.5-Vision using CKA and RSA. The authors report stage- and depth-dependent alignment, correctness-conditioned differences, and advantages over random connectivity. Appendices examine structured nulls, preprocessing, six individual connectomes, a morphological-connectome variant, and estimator sensitivity.

Critical strengths
  1. The separation from readout training is checked directly. Appendices B and M report identical stored reservoir states before and after readout training, with maximum difference zero. This usefully rules out direct modification of those states by the trained readout. The paper also clearly distinguishes FC from anatomical wiring and CWR–VLM alignment from comparison with measured neural activity.

  2. The controls extend beyond uniform-random connectivity. Degree-preserving rewiring, ROI permutation, community and small-world controls, morphological connectivity, and individual reservoirs test different aspects of the construction. The ROI-permutation comparison gives a larger alignment difference than several other nulls (Table 13: approximately 2.1–2.2 times), despite preserving the original matrix's spectrum. The reported debiased-estimator caveats and sign ablations are also informative. Appendix L's six individual matrices provide useful evidence that the qualitative pattern is not confined to group averages.

Critical weaknesses
  1. The retained biological information and the selection procedure need closer examination. The pipeline interpolates the FC matrix, replaces its signs with random signs, and uses a rank-one Kronecker approximation capturing 32.67% of the factorized matrix's reported Frobenius variance (Appendices C and G). Original sign placement gives 50.44% accuracy, versus 61.96% with random signs and 64.70% with sign shuffling (Appendix M). These results establish a trade-off, but do not show whether the alignment advantage survives preserving empirical signs. Table 3 also selects sparsity by the biological-vs-random gap and leak rate by accuracy plus H3 RSA. The selection data and its separation from final evaluation are therefore important to the structural interpretation.

  2. The evaluation leaves uncertainty about effect size, replication, and the interpretation of alignment. Table 13 reports control-specific experiments and explicitly warns against comparing absolute CKA across rows. The uniform-random ratios are 1.080 and 1.095 for Qwen and Phi; the separate N=402 spectrum-matched evaluation gives 1.049 and 1.104, with wins in 5/7 and 6/7 tasks. These are modest relative advantages, whereas the ROI effect is larger. Several task-level intervals cross zero under debiasing. The headline comparisons use fixed group templates across seeds: Appendix L adds individual robustness evidence, but does not estimate a subject-level distribution of empirical-minus-null effects. The reported 63/63 comparisons also need a defined aggregation unit and an uncertainty analysis accounting for shared templates.

    H1/H2 are shared across candidates of the same problem, but that does not make representations from different correctly and incorrectly solved problem sets identical. Alignment also depends on the VLM targets and instance weights. Figure 3 shows different early-stage alignment profiles between conditions, so the “unconfounded” argument needs reconciliation with the actual protocol. Separately, Equation 2 still includes the previous state inside the nonlinear update at alpha=1; this removes explicit leaky blending, while retaining recurrence. It therefore does not isolate the absence of temporal dependence.

Questions for authors
  1. Does correct/error refer to paired candidate substitutions or disjoint sets of solved problems? Are the VLM vectors and item weights fixed between conditions? Please explain the H1/H2 alignment differences in Figure 3.

  2. Can the empirical-minus-null CKA differences be evaluated on a common matched-item set, with raw and debiased estimates and bootstrap intervals? Please define the unit behind 63/63 and distinguish uncertainty over instances, training seeds, null matrices, and subjects.

  3. Which data were used to choose sparsity and leak rate, and were those choices frozen before final evaluation? Do the conclusions persist with preset parameters or selection on a disjoint development task?

  4. Does the alignment advantage survive the existing original-sign condition? A sign-preserving accuracy result alone does not answer this.

  5. Does the pattern persist against the same pretrained VLMs before RAVEN LoRA adaptation? Please also specify the VLM problem/prompt format and whether image-token-only pooling preserves the result.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.